A study by three European universities has shown that access to artificial intelligence drastically reduces people’s ability to recognize their own limits: the willingness to say “I don’t know” fell from 44% to 3%, accuracy was reduced to a third, but confidence doubled. The researchers call it “cognitive surrender”.
Artificial intelligence promises to make us more productive, faster and more efficient. But a new study published in July 2026 reveals a very different side of the coin: far from increasing our abilities, access to AI assistants is eroding our ability to think critically and recognize when we don’t know something.
The study data
Researchers from the University of Milano-Bicocca, the École Normale Supérieure and Sapienza University of Rome designed a revealing experiment. They posed questions about visual details of films – such as the color of a team’s uniform in Bend It Like Beckham – using an AI model (Step 3.5 Flash) that tended to fail precisely on this type of question.
The results were striking:
- Willingness to say “I don’t know”: fell from 44% (without AI) to 3% (with AI available)
- Accuracy in answers: dropped from 27% to 9%
- Confidence in the answer: rose from 30% to 76%
As Valerio Capraro, associate professor at the University of Milano-Bicocca, summarizes: “People got much worse, their accuracy dropped to a third, but they were twice as confident”.
The phenomenon of “cognitive surrender”
The term “cognitive surrender” (cognitive surrender) was coined by Wharton School researchers in early 2026 to describe exactly this pattern: people who accept incorrect AI answers 80% of the time, while reporting higher confidence levels than those who work without it.
The worrying thing is not only that people trust erroneous AI answers. The study shows that the mere availability of artificial intelligence suppresses the cognitive habit of recognizing what one does not know. Even participants who initially answered correctly on their own, upon consulting the AI and receiving a wrong answer, ended up accepting it.
Economic incentives barely help
One of the questions the researchers asked was whether offering economic incentives could reverse the effect. The results were discouraging: the willingness to admit ignorance rose from 3% to 8%, and accuracy from 9% to 16%. Both figures remained far below the levels of the group that worked without AI (44% and 27% respectively).
Implications for education and society
Capraro is especially concerned about children and adolescents: “For humans, the ability to say ‘I don’t know’ is very important because it represents the recognition of the limits of our own knowledge”. Young people are growing up with AI systems integrated into their daily lives, before having fully developed their critical thinking.
The concern extends to the educational field. Common Sense Media recently rated Google’s redesign of its search engine with AI-generated summaries as an “unacceptable risk” for students. The pattern is consistent: AI products are designed to answer, never to say “I don’t know”, and the humans who use them are learning to do the same.
What can we do?
The researchers suggest several lines of action:
- Critical digital education: teach how to evaluate AI answers instead of accepting them passively
- AI design that admits uncertainty: systems should indicate their confidence level in their answers
- Transparency about limitations: AI assistants should explicitly point out what kind of questions they can answer well and which they cannot
- Ongoing research: it is necessary to continue studying how interacting with AI affects our cognitive processes in the long term
The study adds a striking data point to a debate that is just beginning: as artificial intelligence becomes integrated into every aspect of our lives, the question is no longer just what AI can do for us, but what AI is doing to us.
Source: The Next Web / Study by University of Milano-Bicocca, École Normale Supérieure and Sapienza University of Rome. Published July 2026.
